Mathematica tips for numerical linear algebra (3)

Mathematica is smart

The take-away from this is that many clever people have worked for a very long time on optimising the Mathematica’s built-in functions, such as Dot. In particular, when Dot is called, it turns the calculation over to a set of C++ libraries for numerical linear algebra known as the Intel MKL. Of course, we could just do all of this in C++ and call MKL directly. Usually, however, the time saved in the calculation itself is not worth the extra headaches of coding in C++ and having to recompile just to try out a new idea. Mathematica allows us to play with ideas and have performance close to (within a factor of three or so) doing everything in C++. And as we’ll see in a future post, Mathematica can take advantage of multiple CPU cores without us really having to think about it.

We can see the same behaviour play out with higher-rank tensors. Imagine we want to contract a matrix with another matrix: $m_{ab}n_{bc} = (mn)_ac$. We could implement this as

m = RandomComplex[{}, {500, 1000}];
n = RandomComplex[{}, {1000, 750}];

m.n; // AbsoluteTiming
(* 0.0735985 s *)

Not bad! Just let Mathematica do the dirty work! Note that Mathematica thinks of these matrices as arrays (or more accurately Lists) of dimensions {500,1000} and {1000,750} respectively:

m // Dimensions
(* {500,1000} *)
n // Dimensions
(* {1000,750} *)
m.n // Dimensions
(* {500,750} *)

This also gives us an idea of what Mathematica is doing with Dot – the length of the last dimension of $m$ must match the length of the first dimension of $n$ (since it is contracting these indices). The resulting array is labelled by the remaining indices of $m$ and $n$ in the order they appear in according to $m_{ab} n_{bc} = (m n)_{ac}$.

Anthony Ashmore
Anthony Ashmore
Marie Skłodowska-Curie Fellow